Agentic AI a banki yanzu matsala ce ta injiniya da aka rufe ta cikin sutturar matsalar AI. Model ɗin ana iya musanyawa; control plane ba a iya musanyawa ba. Ƙalubalen 2026 ba shi ne shigarwa ba — Cambridge CCAF ya sanya wannan a 52% — ƙalubalen shi ne ko tsarin masu cin gashin kai da bankin ku ke gudanarwa yau zai iya wuce binciken SR 11-7 a kashi uku na watanni masu zuwa. Mafi yawa ba za su iya ba.
Takaitaccen Bayanin Zartarwa / Manyan Abubuwan da za a Koya
- A daina kiransu chatbots. Sashin samar shi ne bounded workflow tare da tsattsauran izini na tool-call. Aikin yana faruwa a cikin workflow, ba a cikin LLM ba.
- OSWorld a 66.3% shi ne rufin dogaro. Benchmark mafi kusa da amfani da tool a manyan masana'antu na Stanford HAI har yanzu yana kasawa a ɗaya daga cikin ayyukan tsararru uku. Wannan lamba ne wanda ke tallafa wa aikin human-in-the-loop mai ƙarfi; ba ya tallafa wa aiwatarwa ba tare da kulawa ba a kan duk abin da ya shafi kuɗin abokin ciniki.
- Rarrabe da izini, ba da hankali ba. Tsanin Cin Gashin Kai yana gudana daga Level 0 (cire sashin ISDA ta hanyar karatu kawai) zuwa Level 4 (gyaran biyan kuɗi na multi-tool tare da tilas checkpoints). Level 5 — aiwatar da kai-tsaye ba tare da checkpoints ba — bai kamata ya kasance a banki na samar a 2026 ba.
- Agent Control Plane sashe biyar ne na injiniya, ba takardar manufofi ba. Service accounts da aka tsara da OAuth, semantic routing mai tabbaci, Open Policy Agent gating, WORM audit logging, da kill switch da aka gwada. Duk abin da ya rasa shi ne binciken.
- SR 11-7 da PRA SS1/23 sun riga sun yi aiki. Fed ya bayyana sau da yawa cewa duk tsarin decisioning na input-zuwa-output yana a cikin iko. Bankunan da suka yi jayayya cewa LLM ba model ba ne sun riga sun yi rashi a muhawarar tsare-tsare kafin su yi.
Me Ya Sa 2026 Shi Ne Shekarar da Wannan Index ya Yi Muhimmanci #
Sauyin daga chat zuwa bounded workflows shi ne kawai abin da ke da muhimmanci a agentic AI ga bankuna a wannan shekarar. Chatbot wanda ke shirya imel ɗin abokin ciniki ana iya bita. Agent wanda ke kira POST /accounts/{id}/freeze a kan dandalin katin samar shaidar audit ce. Samar ya kama tsarin: binciken Cambridge CCAF na 2026 ya ba da rahoton 52% na shigar agentic mai aiki da 23% a matsayi na scaling ko canjawa (Cambridge CCAF ⧉). An wuce iyakar "gwajin keɓe" wajen ƙarshen 2025.
Abubuwa biyu sun canza tare da shigarwa.
Na farko, masu tsare-tsare sun daina ɗaukar LLMs a matsayin sabon abu. Federal Reserve ya bayyana cewa SR 11-7 ⧉ yana aiki kan decisioning na tushen LLM ba tare da la'akari da ko an rarrabe LLM ɗin a ciki a matsayin model ba. SS1/23 ⧉ na PRA koyaushe yana da fa'ida sosai don ɗaukansu. Rarrabe haɗari mai girma na EU AI Act ya rufe yawancin amfani da LLM a sabis na kuɗi. Babu sauran muhawarar "ba mu da tabbaci ko wannan ya shafi".
Na biyu, gaskiyar benchmark ya kama. AI Index na Stanford HAI na 2026 ya ba da rahoton OSWorld — benchmark mafi kusa da ake da shi ga amfani da tool na masana'antu na hakika — a 66.3% daidaito (Stanford HAI ⧉). Ɗaya daga cikin ayyukan tsararru uku har yanzu yana kasawa. Wannan lambar ta sa rufin fasaha kan cin gashin kai a 2026. Mai girma da zai tallafa wa bounded Level-3 deployments a ƙarƙashin kulawar HITL; ba mai girma ba da zai tallafa wa aiwatarwa ba tare da kulawa ba a kan duk API da ke taɓa kuɗin abokin ciniki.
Agentic AI Index don bankuna yana buƙatar yin abin da Basel framework ya yi don capital ga decisioning na tushen LLM: maida ikirari na "muna da sarrafawa" zuwa shaidar da za a iya auna, da binciken audit a kowane workflow.
Tsarin Index na 2026 #
| Layin Index | Yadda "Shirin" Yake | Ma'aunin Shirin | Yanayin Kasawa |
|---|---|---|---|
| Matakin cin gashin kai | Kowane workflow na samar an yi masa lakabin Level 0–4; babu Level 5 a samar | % na workflows ta matakin; rabo a Level 3+ | Agent na samar ya fitar da pacs.008 zuwa beneficiary BIC da aka ƙirƙira saboda babu static allow-list da ke gating payload ɗin kafin SWIFTNet |
| Izinin API | Kowane agent yana taswira zuwa service account guda ɗaya tare da OAuth scopes mai least-privilege (e.g., card-freeze:write:lt-5000usd); MTLS zuwa legacy core |
% na agents a least-privilege; ƙidayar orphan-permission | Agent ya sake amfani da service account mai over-scoped; ya yi maimaitawa kan asusun da ba ya da hakkin karantawa; rahoton GDPR Article 33 incident a cikin sa'o'i 72 |
| Deterministic guardrails | Kowane tool-call ya bi ta semantic router (NeMo Guardrails / LangChain Guardrails) tare da JSON-schema validator kafin API | % na tool-calls da aka katse; ƙimar ƙin yarda ta nau'i | LLM ya fitar da kira na transfer tare da amount: 0; downstream API ba ya tabbatarwa; faɗakarwar daidaita ledger ya iso bayan sa'o'i 18 a yankin lokaci dabam |
| Ɗaukar nauyin human-in-the-loop | Kowane Level-3 execution yana fitar da UI na amincewa tare da tilas timeout; auto-approve an kashe ta manufa | Yawan amincewa; ƙimar rubber-stamp (an amince a cikin ƙasa da daƙiƙa 2) | Operator ya danna "approve" a kan faɗakarwa 200 a cikin minti 4; an gabatar da SAR akan abokin ciniki na gaskiya; korafi na mai tsare-tsare a cikin makon |
| Cikar audit | WORM log mai immutable yana kama system prompt + retrieved context + LLM output + tool-call + tool result + approver UID; an sa hannu cryptographically a lokacin rubutawa | % na invocations tare da cikakkiyar trace | Mai binciken SR 11-7 ya tambaya me ya sa agent #4421 ya amince da wire na $4.8M; bankin yana da rasit ɗin wire da model card; babu shaida ta matakin prompt; an ba da binciken |
| Tattalin arziƙin kowane aiki | An bibiyi kuɗin kowane shawara da aka kammala gami da kuɗin sake mayarwa da gyarawa; tabbatacce vs manual baseline | Net cost na kowane shawara; ƙimar sake mayarwa | Kuɗin kowace token a kan agents na shari'ar gefe ya wuce kuɗin manual investigator da suka maye gurbi; CFO ya kashe shirin a Q3 |
Alamomi na Yanzu da za a Bibiya #
| Alama | Abin Da Yake Nufi ga Bankuna | Tushe |
|---|---|---|
| 52% shigar mai aiki | Agentic AI ya wuce matakin gwaji; mulki na cibiya gabaɗaya ya makara | Cambridge CCAF ⧉ |
| 23% scaling ko canjawa | Ɗan tsiraru mai muhimmanci ya wuce gidan wasan kwaikwayo na tabbatar da fahimta | Cambridge CCAF ⧉ |
| OSWorld a 66.3% | Ƙimar kasawa ɗaya-cikin-uku akan amfani da tool tsararru. Aiwatarwa ba tare da kulawa ba a kan API ɗin kuɗin abokin ciniki ba ya tallafawa a wannan matakin dogaro | Stanford HAI ⧉ |
| 55% sun yi nuni da rasa kulawar mutum a matsayin babban haɗari | Ƙirar sarrafawa ita ce babbar damuwar injiniya, ba na compliance na ƙarshe ba | Cambridge CCAF ⧉ |
| 76% na manyan FIs sun yi gwagwarmaya don auna daraja | Ikirarin yawan aiki na gabaɗaya ba ya rayuwa a tattaunawar CFO. Auna kowane workflow, ba kowane shiri ba | Cambridge CCAF ⧉ |
Tsanin Cin Gashin Kai #
Rarrabe agents bisa abin da aka ba su izinin yi, ba bisa hankalin model ɗin da ke ƙasa ba. Hakika GPT-5 / Claude 4 / Gemini 3 zai iya zama a kowane matakin; wrapper shi ne abin da ya bambanta.
- Level 0 — Lura. Karatu kawai zuwa logs, traces, ko ma'amaloli. Agent yana fitar da tsare-tsare ko abubuwan da ba na al'ada ba; babu rubutawa ko'ina. Misali: gano drift a
pacs.008rejection rates ta corridor da faɗakar da ƙungiyar ayyuka. - Level 1 — Karatu kawai. Yana karantawa daga tsarin ayyuka; yana fitar da tsararriyar output don amfanin ɗan adam. Misali: cire bambance-bambancen sashin CSA daga ISDA Master Agreement na counterparty da nuna keɓancewar daga template na ma'auni na bankin. Agent ba ya rubuta dawowa ga shagon kwantiragi.
- Level 2 — Shirya don ɗan adam don shigarwa. Yana samar da abun ciki da ɗan adam ke bita da gabatarwa. Misali: shirya Suspicious Activity Report daga faɗakarwar tsarin zamba gami da rikodin KYC gami da transaction trace; jami'in BSA yana karantawa, gyarawa idan an buƙata, da gabatarwa. Tsarin rikodin yana ganin sigar da ɗan adam ya amince da ita kawai.
- Level 3 — Aiwatar da bounded. Yana kiran API na samar tare da tsattsauran iyakar deterministic da wrapper ke tilastawa. Misali: kira na API na card-freeze tare da
max-amount-at-risk: 5000 USDda allow-list policy ke tilastawa; agent ba zai iya daskare katin da aka haɗa shi da balances sama da wannan iyaka ba tare da escalation na Level-2 ba. Iyakar tana zaune a policy-as-code, ba a cikin prompt ba — prompts ba security boundary ba ne. - Level 4 — Multi-tool orchestration tare da tilas checkpoints. Yana gudanar da jeri a tsarin; an yi log na kowane sauyi na yanayi; checkpoints suna buƙatar amincewar ɗan adam kafin tool-call na gaba. Misali: workflow na gyaran biyan kuɗi — cire
pacs.008da ya kasa daga dead-letter queue → nemo daidai beneficiary ta SWIFT KYC Registry → samar da saƙon da aka gyara → rubuta zuwa outbound queue → ɗan adam ya amince da sake aikawa. Idan kowane mataki ya kasa schema validator, workflow yana tsayawa kuma yana ƙirƙirar shari'ar keɓancewa. - Level 5 — Self-orchestration. Agent yana tsarawa kuma yana aiwatarwa ba tare da amincewar checkpoint ba. Babu workflow na banki na samar wanda ya kamata ya kasance a Level 5 a 2026. Wannan ba bayanin tsufa ba ne; bayanin dogaro ne. OSWorld a 66.3% yana tarawa a tsakanin haɗin kai na API. Tool-calls uku a 66% kowane shi ne 29% na nasarar ƙarshe-zuwa-ƙarshe. Biyar shi ne 13%. Kar a yi.
Agent Control Plane #
Control plane shi ne layin injiniya tsakanin LLM da tsarin samar. Sashe biyar, duk runtime, babu wanda aka rubuta a cikin takardar manufa.
1. Asali da Izini #
Kowane agent yana taswira zuwa service account guda ɗaya kawai. Wannan account yana riƙe OAuth client_credentials tokens da aka tsara zuwa mafi ƙanƙantar fuskar API da ake buƙata. Token na card-freeze agent zai iya kira POST /accounts/{id}/freeze tare da amount-at-risk: 0..5000 usd. Ba zai iya kira GET /accounts/{id}/balance ga sauran abokan ciniki ba. Ba zai iya kira komai a custody, treasury, ko trading ba. Asirin service-account suna jujjuyawa kowane mako; long-lived credentials shi ne kasawar control-plane mafi yawan gani a deployments na samar.
2. Deterministic Guardrails a kan Tool-Calls #
Kowane LLM tool-call yana bi ta deterministic semantic router (NeMo Guardrails, LangChain Guardrails, ko makamancinsa) kafin kiran ya isa API na samar. Router yana rarrabe niyya akan ƙayyadaddiyar allow-list; ana ƙin kira a wajen jerin kuma a yi log. Sa'an nan JSON-schema validator yana duba payload — fage masu buƙata sun kasance, adadin dollar a cikin iyaka, lambobin ƙasa ISO masu inganci, beneficiary BIC a kan jerin counterparty na bankin da aka riga aka amince. Validator ya kamata ya kasance da yawan tuhuma: pacs.008 tare da amount: 0 kasawar model ce, ba ma'amala ta gaskiya ba. Haka kuma wire zuwa ƙasar da sanctions filter ɗinku bai riga ya amince ba ga kashin abokin ciniki na asali.
3. Policy-as-Code #
Open Policy Agent (ko makamancinsa) yana zaune tsakanin validator da API. Manufofi suna da sigogi a Git; shawara na ƙin yarda ana yin log; ainihin policy engine wanda ke gating kiran microservice-zuwa-microservice a dandalin ku na yanzu yana gating tool-calls na agent. Ɗaukar agents a matsayin keɓaɓɓen aji tare da bespoke gating shi ne yadda bankuna ke ƙarewa da control planes na inuwa wanda babu wani a tawagar dandalin ke fahimta bayan watanni shida.
4. Audit Logging #
Ajiyar WORM mara canjawa — S3 Object Lock, Azure Blob immutability, ko ledgered database. Kowane invocation yana kama: timestamp, agent ID, service-account ID, hash na system-prompt, retrieved context, mai bayar da LLM gami da model gami da version, raw LLM output, parsed tool-call, shawarar OPA, amsar API, sakamako na downstream, da approver UID inda ya dace. Rikodin an sa hannu cryptographically a lokacin rubutawa. Wannan log shi ne abin da masu bincikar SR 11-7 da SS1/23 za su nema. Idan ba za ku iya samar da cikakkiyar trace don kowace shawara ba, ba ku da agent da aka gudanar da haɗari na model.
5. Kill Switch #
API mai red-button wanda ke soke duk invocations na agent a cikin ɗimbin izini a cikin ƙasa da daƙiƙa 60. An gwada kowane kashi uku na watanni tare da tabletop exercise. Kill switch shi ne kawai abin da yake murmurewa ku daga sakin model na mai siyarwa wanda ke koma baya cikin shiru, vector na prompt-injection wanda ba ku tsammani ba, ko taron drift wanda ke turawa ƙimar false-positive ta wuce iyakar ayyukanku. Kill switches da ba a gwada ba ba sa aiki; tsara lokacin motsa jiki.
Gudanar da Haɗari na Model #
Bankunan da suka yi jayayya "LLM ba model ba ne a ƙarƙashin SR 11-7" sun riga sun rasa. Federal Reserve ya bayyana sau da yawa cewa duk tsarin input-zuwa-output da ake amfani da shi a workflow na decisioning yana a cikin iko. PRA's SS1/23 ya fi fa'ida. Daidaitaccen matsayi: ɗauki kowane agent na samar a matsayin model na SR 11-7 / SS1/23 daga rana ta farko. Kuɗin sake tsara agent da aka tura a matsayin model shi ne sau da yawa kuɗin tsara shi haka tun da farko.
Layuka uku na tsaro, an yi amfani da su a agents:
- Layi na farko (mai mallakar model). Yana rubuta amfani da agent da aka nufa, lineage na training da eval data, schema na system prompt, allow-list na tool-call, sakamakon gwajin kill-switch. Yana mallakar sa idanu na drift a samar.
- Layi na biyu (ƙungiyar MRM). Yana tabbatar da agent kafin samar. Rahoton tabbaci ya rufe maki na eval da mai siyarwa ya saki (MMLU, HumanEval, HellaSwag suna da amfani amma ba isasshe ba), maki na eval na musamman ga bankin (saitin evaluation na ku da aka gina daga misalan ayyuka — wannan shi ne aikin da yawancin bankuna suka rage zuba jari), sakamakon red-team na prompt-injection, nazarin son zuciya da adalci inda workflow ke da tasiri ga abokin ciniki, da bayanin lissafi na residual-risk.
- Layi na uku (audit na ciki). Yana gwada gates na control-plane da cikar audit-log akan samfurin shawarwarin samar. Tsarin audit na 2027 zai yi kama da na 2025 sosai; tsara kasafin shi yanzu.
Sa idanu na ci gaba ya fi muhimmanci fiye da tabbaci na lokaci ɗaya. Bank-specific eval suites da ake sake gudanarwa kowane mako suna kama regressions na sabuntawar model wanda vendor benchmarks ba za su fito da su ba. Yawan sakin na OpenAI, Anthropic, da Google ya fi yawan tabbatarku; ko kuma za ku rufe gibin ta hanyar gudanar da continuous evals, ko kuma binciken mai bincike zai rufe shi a gare ku.
Auna Tasirin Kasuwanci #
Ikirarin yawan aiki na gabaɗaya ba ya rayuwa a tattaunawar CFO. Auna agents kamar yadda kuke auna sauran sauye-sauye na aiki:
- Kuɗi a kowace shawara da aka kammala, gami da kuɗin sake mayarwa da gyarawa na shawarwari da suka kasa. Agent na shirya SAR wanda ya yanke lokacin jami'in BSA da 40% amma ya samar da 12% na false-positive filings ya lalata daraja, ba ya halicce ta ba.
- Manual touches da aka kauce wa, an ƙidaya net na sabbin touches da kulawar control-plane da kula da keɓancewa ya halitta. Manufa ba ita ce a rage kulawar ɗan adam ba; ita ce a juya shi zuwa shawarwari masu girma.
- Ƙimar sake mayarwa — kashi na ayyukan agent-executed da aka mayar da su a cikin sa'o'i 24. Ƙimar sake mayarwa sama da 2% akan workflow na Level-3 matsala ce ta dogaro. Sama da 5% matsala ce ta control-plane.
- Cikar audit-trace — kashi na shawarwari tare da cikakken provenance da za a iya sake gina shi daga WORM log. Ya kamata ya kasance 100% akan workflows na Level-3 da Level-4. Komai ƙasa kasawar manufa ce wanda zai bayyana a audit.
Idan workflow ya zama mafi sauri amma ƙasa da bayyananne, index yana buƙatar hukunta shi. Hanya mafi arha don kasa jarrabawar mai tsare-tsare ita ce inganta yawan aiki da rasa trace.
Abin Da Wannan Ke Nufi ta Nau'in Banki #
Global Systemically Important Banks #
Matsalar wuya ita ce mulki a girma: ɗaruruwan agents a kan layukan kasuwanci, kowane yana da mai mallakar model nasa, kowane yana iya zama binciken audit. Zuba jari ba sabon gwaji ba ne. Shi ne central control plane, hadaddiyar audit-log infrastructure, da MRM bench mai ikon tabbatar da agents 50-da ƙari a kowane kashi uku na watanni. Ba tare da wannan iyawa ba, agents suna sauka da sauri fiye da yadda za a iya gudanar da su kuma cibiyar tana tarawa fitarwar SR 11-7 cikin shiru.
Transaction da Corporate Banks #
Workflows masu mafi girman ROI sune gyaran biyan kuɗi, cire takardun KYC, kauce wa FAQ na treasury-services, da rabe-raben daidaituwa. Duk Level-2 ko bounded Level-3. Abokin ciniki na corporate ba ya damuwa da cewa agent ne ya yi aikin; suna damuwa cewa SLA ya inganta kuma ƙimar jayayya ta tsaya tsayin daka. Shiga gaba tare da ma'aunin, ba fasaha ba.
Bankunan Yanki #
Saya, ba a gina ba. Ɗauki vendor wanda dandalin agent ɗinsa ya riga ya sami primitives na control-plane — OAuth scoping, OPA integration, WORM audit logging, kill switch da aka gwada — kuma tabbatar da wannan dandalin akan MRM framework ɗinku. Gina bespoke control plane shi ne zuba jari na shekaru da yawa wanda ba ya bambanta a girman yanki. Kashe ƙarfin injiniya akan ƙirar workflow da operator UX maimakon.
Fintechs, PSPs, da Masu Bayar da Infrastructure #
Tambayar samfurin ga masu siyarwa ba "shin AI agent ɗinku yana yin aiki mafi kyau fiye da mutane" ba. Ita ce "shin dandalin ku yana samar da audit trace mai bin SR 11-7 daga akwati." Masu siyarwa waɗanda za su iya amsa hakan da i za su rufe yarjejeniyoyin masana'antu. Masu siyarwa waɗanda ba za su iya ba za su makale a cikin maimaitawar tabbatar da fahimta yayin da ƙungiyar MRM ta bankin ke nemo dalilai don kasa tabbatar.
Ƙarshe #
Agentic AI a bankuna a 2026 matsala ce ta injiniya. Aikin mai ban sha'awa yana cikin control plane, ba a cikin model ba. Model ɗin ana iya musanyawa; OAuth scoping, deterministic semantic router, OPA policy gates, audit log mai immutable, da kill switch ba a iya musanyawa ba.
Cibiyoyin da za su yi kama da abin yarda ga masu tsare-tsare a cikin watanni 18 sune wadanda ke ɗaukar kowane agent na samar a matsayin model na SR 11-7 / SS1/23 daga rana ta farko, tare da bank-specific eval suites na ci gaba da kill switch da aka gwada, da control plane da aka gina don kasawa cikin aminci. Cibiyoyin da ba sa yi za su gano ko MRM bench ɗinsu zai iya girma don sarrafa fiye da 50 na bincikan gyara kowane kashi uku na watanni.
Auna agents kamar yadda kuke auna kowane sauyi na aiki: kuɗi, dogaro, ikon mayarwa, shaida. OSWorld a 66.3% shi ne rufin dogaron ku. Tsara daidai.
Tambayoyin Da Aka Saba Yi #
Mene ne agentic AI a banki?
Bounded workflow wanda ke haɗa LLM tare da tool-calls a cikin tsarin samar, runtime guardrails, da checkpoints na human-in-the-loop. Aikin yana faruwa a cikin workflow, ba a cikin model ba. Idan kun ji kalmar "chatbot", kuna a cikin nau'in da bai dace ba.
A ina ya kamata bankuna su fara?
Workflows na Level 1 da Level 2 inda za a iya auna daraja kuma ana iya rufe ɓangaren mara kyau: cire sashin ISDA, shirya SAR, triage na gyaran biyan kuɗi, dawo da ilimi na ciki, taimakon bita na code, rarraba takardar KYC. Tsallake Level 3 har sai control plane ɗinku ya sarrafa OAuth scoping, semantic routing, OPA gating, WORM logging, da kill switch da aka gwada.
Mene ne babban haɗari?
Barin agents su aiwatar akan APIs na samar ba tare da deterministic guardrails tsakanin LLM output da API ba. Lambar OSWorld 66.3% ita ce gargaɗi. Tool-calls da ba a nannade ba a wannan ƙimar kasawa akan SWIFT MT103 ko API na kuɗin abokin ciniki za su rubuta mafi muni a kanun labarai na zagayowar tsare-tsare na gaba.
Shin SR 11-7 yana shafar agents na tushen LLM?
Eh. Federal Reserve ya bayyana cewa duk tsarin input-zuwa-output da ake amfani da shi a decisioning workflows yana ƙarƙashin SR 11-7. SS1/23 na PRA ya rufe ƙasar ɗaya a UK. Rarrabe haɗari mai girma na EU AI Act ya rufe yawancin amfani da sabis na kuɗi. Muhawarar "shin wannan model ne" ta ƙare; yi aiki daidai.
Yadda ya kamata a ba da rahoton agentic AI ga allunan gudanarwa?
Lambobi huɗu a kowane workflow: matakin cin gashin kai, cikar audit-trace, ƙimar sake mayarwa, net cost na kowace shawara. Da kuma jerin manyan biyar na residual-risk. Tsallake model-card slideware.
Tushe #
- Stanford HAI, (2026). Rahoton AI Index na 2026 ⧉.
- Stanford HAI, (2026). Babi na Aikin Fasaha ⧉.
- Cambridge Centre for Alternative Finance, (2026). Rahoton AI a Sabis na Kuɗi na Duniya 2026 ⧉.
- Federal Reserve, (2011). SR 11-7: Jagora kan Gudanar da Haɗari na Model ⧉.
- Prudential Regulation Authority, (2023). Supervisory Statement SS1/23: Ƙa'idodin gudanar da haɗari na model ga bankuna ⧉.
- European Commission, (2024). Regulation (EU) 2024/1689 — AI Act ⧉.
- NVIDIA, (2024). Framework na NeMo Guardrails ⧉.
- Cloud Native Computing Foundation, (2018). Open Policy Agent (OPA) ⧉.
Bita ta ƙarshe .
Bita ta ƙarshe .